Skip to content

04 — Processes

Status: Current | Last reviewed: 2026-10-20

How the systems documented in 03-services/ cooperate to do specific things. Read these when you are about to change one.

  • Analytics Agents — native transcript preparation, configurable canvas, private evidence chat and migration boundaries.
  • Analytics Agent types and canvas — Call Drivers, type-specific configuration and preset steps.
  • Analytics dataset inputs — operational table imports, version carry-forward and large intent hierarchies.
  • Analytics composition — selected component catalogues, analytical tool policies, prompts and briefing schedules.
  • Universal pipeline builder — how an agent's JSON config becomes a running pipeline. Read first.
  • OpenAI GPT Live — continuous voice, configurable delegation, tools, multimodal input, and usage.
  • Voice startup preparation — overlap caller setup with welcome audio and join the same durable workflow.
  • Consumer Cellular demo — ANI greeting, sample AutoPay confirmation, and friendly caller-led conversation.
  • Consumer Cellular native Connect demo — human-agent assistance, native Cases, knowledge, notes, phone routing and rollback.
  • Demo qualification and city parity — saved Studio cases, repeated strict verdicts, Copilot replay, and Vanguard/Macau validation.
  • ElevenLabs voices and half-cascade agents — v4 model discovery, expressive formatting and streaming text with a separate voice.
  • SIP integrations — Genesys, Webex, Five9, Amazon Connect, Salesforce SCV.
  • Real-time translation — translation pipeline subtype.
  • AWS real-time translation — two Polly streaming modes, SCV integration and repeatable measurements.
  • AI counterparty detection — continuous Voice/Chat and Copilot assessments, live variables, cards and OpenMuse evaluation.
  • Guided channel selection — proposed shared channel eligibility, queue diversion, callbacks, acoustic signals and OpenMuse voice testing.
  • Alder Homeworks guided-service storyboard — draft scripts, fact continuity, spoken web handoff, owner approval and video-inspection production.
  • Copilot API — a TrueServe API in front of the agent-assist pipeline for custom LWCs, with a swappable AI backend.
  • DevSDO Lightning UI audit — live app branding, three-column page compositions, component catalog and shared-toolkit candidates.
  • Live copilot ingress — select a copilot, attach or stream audio once, and share the resulting session across page widgets.
  • Takeda demo — Voice Agent self-service, Renee's handoff to the human Copilot, dynamic center workspace, measured transcription latency and the DevSDO app.
  • Takeda call action logic — each Genesys MCP operation, configuration ownership, and requirements for moving the logic into workflows.
  • Takeda native workflows — published Salesforce mappings, action logic, completion rules and migration qualification.
  • Takeda Salesforce mappings and audit — Case, Task, summary and handoff mappings, schema checks, and confirmed completion defects.
  • Surgery Partners patient access — outbound verification, card review, human-controlled patient screens, Salesforce callback receipts and repeatable Preview rehearsals.
  • Allianz phone and Salesforce console — recorded IVR, beneficiary Copilot, producer voice agent and Erik's Omni specialist queue.
  • Allianz DCC human service — Script v2 Copilot events, native form capture, receipt and wrap-up in Support Services.
  • Amazon GREF facilities — separate voice/Copilot agents and the Amazon-branded DevSDO app.
  • Allianz Producer Services Scene — the combined phone and Salesforce journey, saved voices, timed form capture and replay.
  • Warm hand-off — consult-style transfers.
  • Human Supervision — the supervisor copilot loop: one deterministic conversation per call, silent transcript monitoring ([[SILENT]] gate), async Q&A, call-end expiry, post-call resume.
  • Native safety profiles — Llama Guard/NeMo adapters, versioned tenant profiles, response/tool/workflow enforcement.
  • Agent logic — the shared safe expression language behind handoff rules, trigger conditions, and the Agent Logic authoring feature.
  • Customer authentication — verifying the caller's identity: OTP, passkey via multimodal link, tool-based KBA, and the once-per-conversation state model.
  • Prompt auto-injections — the date/time preamble, tenant global instructions, and capability guidance every prompt ships with, and the per-stage overrides in the editor's Auto-injected drawer.
  • Payments (PCI) — taking a card or bank payment without cardholder data reaching a model: the two gates, the tenant-bounds composition, and the isolated capture service.
  • Audio ducking — mixing rules in media-proxy.
  • LangGraph construction — async-agent's workflow runtime.
  • Workflow triggers — what fires async work.
  • Start Conversation node — how a workflow hands a person to an agent: the channel discriminator (chat / SMS / email / outbound voice) and the External-Consult authorization path voice reuses.
  • ai-manager file layout — where to add UI for new pipeline features.
  • Preview Mock Caller — adaptive ElevenLabs caller, private nudges, wrap-up and direct audio recording.
  • Personal assistant Scenes — phone assistant, IVR audio, linked browser chat and Salesforce Messaging Sessions.
  • Director — scripted voice demos, scene timelines, phone and AI Activity playback, and MP4 export.
  • Alder Homeworks Scenes — guided channels, IVR diversion, generated inspection video and explicit owner purchase confirmation.
  • Agent editor errors and warnings — friendly labels, diagnostic inventory, and expandable details.
  • AI Manager prototype experience audit — detailed comparison of the root HTML prototype and current UI, including typography, canvas controls, dropdowns, and interaction gaps.
  • AI Manager experience alignment — implementation decisions and verification for the real app and independent design export.
  • Studio UI prototype — combined conversational design workspace and original editors.
  • Studio dashboard integration — brand/layout decisions, capability preservation and backend follow-up.
  • Studio UI rollout — real-app Home, unified repository, favorites, recents and access checks.
  • Design prototype repository — independent Git submodule, designer freedom, advisory commit-time impact notes and updates from the framework.
  • Conversational authoring research spike — experience guidance, capability discovery and branded catalog generation (draft).
  • Conversational authoring implementation research — four agent types, tool placeholders, multimodal contracts, source research, and evaluation evidence (draft).
  • Phone emulator — the shared iPhone UI behind Preview, /emulator, and the embed: shadow-DOM isolation, host-served artwork, motion, and the opt-in lock screen.
  • Secret storage — encrypted credentials for tools and integrations.
  • AI generation — the "Generate with AI" creation path and generation MCP.
  • OpenAI authoring — managed Agents API integration, provider switching, recovery, cleanup and Vanguard comparison.
  • Authoring benchmark — 11 September 2026 — latest-model cost, latency, reasoning and draft-quality measurements.
  • Editor assistant — the in-editor Assistant chat for agents/workflows (questions, debugging, live edits).
  • Conversational design — live conversation-led draft creation using the selected authoring engine.
  • Multi-agent demos — coordinated Voice, supervisor Chat and human-to-human Copilot design, build and preview.
  • External Studio management API, MCP, and A2A — OAuth, scoped REST/MCP/A2A authoring, Generate with AI, assistant, diagnostics, and coding-tool skills.
  • Scenario generation skill — reusable voice/Copilot/supervisor build guidance, deterministic test planning, and the API extension plan.
  • A2UI rendering — components-as-tools.
  • Human-controlled customer components — shared send/response API, linked conversations and saved screen snapshots.
  • Record-backed Agent Assist — long-lived CRM rail content without an active conversation.
  • Pharma email operations — SES and Salesforce Email-to-Case intake, workflow policy, durable analysis, and human approval.
  • Agent code export — Strands/Genkit/ADK code-package exports, security policy, and degradation rules.
  • Workflow code export — the LangGraph Python package export: node disposition, vendored kernel, suspend/resume semantics.
  • Presence and collaboration — collab-gateway and the multi-cursor surfaces.
  • Capability packs — the extensible tenant-scoped capability registry: publish-time expansion + library UI (built), AI-assisted authoring and signed industry packs (design).
  • Cost estimation and usage attribution — versioned prices and assumptions, agent/workflow forecasts, durable observed usage, coverage, and reconciliation.
  • Layered PII/PHI protection — design (nothing built yet): five layers keeping PII/PHI out of providers, egress paths, and our own stores while preserving S2S latency.
  • Runtime profiles — design (nothing built yet): one agent, multiple pipeline variants (S2S / cascade / backup / chat), automatic selection and mid-call failover, stack-first authoring.
  • Scripted subagents — deterministic step-machine subagents (STT + TTS + logic/patterns/SLM) that deep-pause the S2S model, mixed voice/DTMF/form capture kept in sync, stack-derived auxiliary services.

  • Procedure supervision + SLM extraction — enforced checklists over a free-running agent: state as session variables, deterministic tool gates on irreversible steps, a trailing SLM monitor with a correction ladder, per-turn entity extraction, and the logic_slm aux slot.

  • Takeda PA copilot — one coaching node, two demo views, and Genesys mapping permissions.

See also